A small town needs occasional multi-day GPU training runs but cannot justify owning a large GPU fleet year-round. Which AI infrastructure type best fits that need?

Cloud AI infrastructure with elastic burst training capacity fits a small town that needs occasional multi-day GPU training but cannot justify owning a large year-round GPU fleet. The correct choice works because cloud elasticity lets the municipality rent accelerators for training windows, then release capacity when idle—matching sporadic training demand without capital lock-in. When residency is not the blocker stated in the stem, burst cloud training is the economical infrastructure type.

Cloud AI infrastructure that provides stateless function training capacity fails because stateless functions target short-lived event-driven executions with strict duration and memory limits, not persistent multi-GPU sessions for occasional multi-day GPU training runs. Cloud AI infrastructure that provides burstable notebook training capacity fails conceptually because notebooks sized for interactive bursty exploration cannot sustain guaranteed multi-GPU throughput across occasional multi-day training windows. Cloud AI infrastructure that provides preemptible spot training capacity fails because preemptible capacity can be reclaimed mid-run, which threatens completion of occasional multi-day runs that require reliable capacity hold, unlike guaranteed elastic burst training.

Exam caveat: cloud burst still requires data-handling review—if later datasets become residency-bound, hybrid or on-prem paths may become mandatory even if elasticity remains attractive. Operational check: estimate annual GPU-hours, compare cloud burst cost to owned fleet TCO, document which datasets may leave the civic boundary, and reject stateless function, burstable notebook, or preemptible spot training options when the requirement is intermittent heavy training.